The blog post

I can't believe I haven't read this until now. This is mind-provoking, and the result is an important step towards understanding neural networks.

Paper

TLDR: Propose ActAdd, a method for controlling model behavior during inference by modifying activations with a bias term that is learned from a pair of prompt.

Summary:

• Propose ActAdd, a method for controlling model behavior by modifying activations at inference time.
• Steering vectors are computed by taking the activation differences that result from pairs of prompts. The vectors are added as bias during inference.
• ActAdd provides control over high-level properties of the output, and preserves off-target model performance, and requires little computational and implementational costs.

I will be holding a seminar at ModelBest (面壁智能) in Sep 20, 2023 in Beijing, Haidian, 科技园. The seminar will be in Chinese, and it's called "大模型安全与伦理问题" (translation: Safety and Ethical Concerns of Large Language Models). Below is a list of references.

Code | Paper (on hold by ArXiv) | Paper (preprints.org) | 知乎

I did this work with my girlfriend, whose research direction is computational fluid dynamics (CFD). We observed that there are numerous research works in applying deep learning (DL) to solve CFD problems. E.g., Pangu-Weather have shown that DL methods can not only be more accurate than the best numerical methods, but can also be multiple magnitudes faster.